Deep Constrained Energy Minimization for Hyperspectral Target Detection

Abstract

Hyperspectral images contain abundant spectral information, which provides great potential for detecting targets that cannot be analyzed with color images. However, a variety of factors, including inherent spectral variability and noise, make it difficult for traditional detectors to separate the target and background by using linear decision boundaries. In this work, we propose a nonlinear detector formulation by generalizing the conventional constrained energy minimization (CEM) method and then design novel nonlinear detectors with two deep neural network structures (named deep CEM or DCEM). The pixel-based structure confirms the effectiveness of the proposed framework, and the cube-based structure utilizing spatial information further improves the performance of the algorithm. The experimental results show that the proposed DCEM method outperforms other competing hyperspectral target detection algorithms.

Keywords

• Deep constrained energy minimization (CEM)
• hyperspectral target detection
• multiple priori target spectra
• nonlinear
• spectral variability